Research On Multi-parameter Joint Inversion Of Shallow Explosion Source Location Based On Deep Reinforcement Learning

As a special application scenario of seismic positioning,the seismic source locating of shallow underground explosions plays a key role in the research of underground damage evaluation,explosion damage effectiveness evaluation and deep penetration weapon test analysis.However,in the actual shallow explosion positioning process,due to the high pulse characteristics of the explosion wave excitation process and the anisotropy of the explosion wave propagation process,it is difficult to pick up the information of the explosion wave arrival time and the nonlinear characteristics of the observation data.Although the traditional positioning method based on waveform information can also locate the observation data with missing information or with multi-medium velocity field,but in the case of unknown parameters of medium velocity field,the fast and high-precision positioning of shallow explosive source still needs to be further solved.In response to these problems,based on the theory of machine learning and multi-parameter joint inversion,this paper focuses on the actual scene of near-field shallow explosion source positioning,and proposes a shallow explosion source scanning positioning method based on deep reinforcement learning method,and a multi-parameter joint inversion source positioning method based on deep reinforcement learning.This method can accurately and quickly locate the source location when the arrival time information is missing and the medium velocity field parameters are unknown.Finally,the results of static explosion tests on a small site(100m*100m*50m)show that the shallow blast source scanning positioning method based on deep reinforcement learning in this paper improves the positioning accuracy(<1m)in the nearfield blast source positioning.On the basis of obtaining the precise positioning of the multiparameter joint inversion source positioning method based on deep reinforcement learning,the velocity field reconstruction with an accuracy of 0.5m is completed within the region,which further improves the interpretability and positioning accuracy of the source location using this method.In this paper,based on the full investigation of the domestic and foreign research status of explosion monitoring and positioning technology and the theory of underground velocity field inversion,according to the actual characteristics and environmental requirements of near-field shallow explosions,the feasibility analysis of using deep reinforcement learning to locate the source is carried out.Aiming at the problem of the lack of original observation data and the mismatch of information dimensions,the three-dimensional scanning source imaging method is used to map the original one-dimensional source data to the three-dimensional energy domain,which solves the problem of deep reinforcement learning to manifest the source characteristics of the input data.At the same time,a fast three-dimensional scanning source imaging method is proposed.By deriving the principle of the existing source scanning algorithm and improving the algorithm,the scanning logic of the scanning algorithm is optimized.In view of the arrangement characteristics of the near-field detector array,the key area brightness attention factor is introduced,which greatly improves the efficiency of the scanning algorithm on the basis of ensuring the completeness of the scanning data,and provides the possibility of expanding the source scanning algorithm in three-dimensional space.On the basis of completing the three-dimensional source imaging,for the nonlinear problem in the iterative process,this paper proposes a shallow source scanning positioning method based on deep reinforcement learning.This method follows a gradual approaching strategy.It associates the guessing point with the three-dimensional source imaging data according to the feature map.The source detection process is regarded as a Markov process,and the source center search strategy is learned through deep reinforcement learning.This method uses deep reinforcement learning to perform characteristic fitting on nonlinear problems,and improves the accuracy and efficiency of the explosion source location.In order to further deepen the interpretability and positioning accuracy of reinforcement learning in the location of shallow explosion sources,in view of the impact of underground medium velocity models on energy imaging during data inversion,this paper proposes a multi-parameter joint reconstruction method of source location and medium velocity model based on reinforcement learning.This method introduces a three-dimensional source imaging process in the multi-parameter joint inversion process to strengthen the correspondence between the energy domain information and the medium velocity model in the forward process.Then,in the inversion process,the energy field of the waveform is inverted by the known source information and the medium velocity field model.On the basis of obtaining the wavefield forward waveform and inversion waveform,the update process of the velocity model is regarded as a Markov process,the correspondence between the forward model and the velocity model is established,and the speed model is gradually trained and optimized through deep reinforcement learning to obtain the shallow source location results under multi-media parameters.In the end,this method can complete high-resolution medium velocity field reconstruction and source location without the medium parameters being known.